AMD launches Helios rack-scale AI platform and broader compute portfolio
At its Advancing AI 2026 event, AMD introduced Helios, its first rack-scale AI system, pairing 72 Instinct MI455X GPUs with 18 sixth-generation EPYC "Venice" processors and Pensando networking. The company said Helios delivers up to 30% more inference tokens per dollar than competing rack-scale solutions and disclosed working relationships with OpenAI, Anthropic, Meta, Microsoft, Oracle, and Supermicro to deploy or support Helios-based infrastructure. AMD also rolled out the wider MI400 GPU family, including the MI455X for AI training and inference, the MI430X for scientific computing and sovereign AI, and the MI350P for acceleration in existing systems, alongside the ROCm.ai development platform that supports PyTorch, Hugging Face, vLLM, SGLang, and coding assistants such as Claude, Codex, and Cursor. New Ryzen AI Embedded X100 processors and Kria AI modules extend the lineup into edge and robotics, with AMD describing a compute roadmap stretching through 2030.
Majestic Labs unveils Prometheus server built around unified LPDDR6 memory
Tel Aviv-based startup Majestic Labs introduced the Prometheus server, which replaces Nvidia GPUs and high-bandwidth memory with up to 12 custom "Ignite" AI Processing Units that combine Arm cores with RISC-V vector and tensor engines, backed by a pool of 8TB to 128TB of coherent LPDDR6 memory per system. The company claims the design delivers more than 50 times the fast memory of an Nvidia DGX B300 configuration and 1.7 times the interconnect bandwidth, with custom memory aggregation chiplets linked by copper cables up to one meter long. A standard 40U rack holds four Prometheus servers drawing 120 kilowatts and relying on cold-plate liquid cooling rather than forced air, and the system is built to Open Compute Project specifications with native support for PyTorch, vLLM, and OpenAI's Triton framework. Majestic Labs, founded in 2023 by CEO Ofer Shacham, President Sha Rabii, and COO Masumi Reynders and staffed by roughly 40 employees across Tel Aviv and Los Angeles, closed a $100 million Series A in late 2025 and is targeting shipments next year with projected pricing 10 to 50 times below equivalent GPU-based systems, though independent benchmarks have yet to appear.
OpenAI advances a $20 billion data center campus in coastal Georgia
Reporting from the Atlanta Journal-Constitution detailed how OpenAI's planned 4.4 million-square-foot campus in Effingham County, announced July 22, took shape under nondisclosure agreements that bound local officials through the site selection and tax negotiations. Under a 15-year payment-in-lieu-of-taxes arrangement, the company will pay 50% of the property's appraised value at the 2026 tax rate each year and is committed to hiring a minimum of 400 workers at an average wage of $80,000, with no mandate for benefits and the option to use third-party contractors. The pivotal infrastructure link was Georgia Power's Plant McIntosh, which will produce 2.1 gigawatts once new turbines come online and through which the utility's economic development arm introduced OpenAI to the county's industrial development authority last fall. The article traces how the project moved through a February 2025 commissioners retreat where data centers were mentioned only as part of a Georgia Power presentation before being kept off the public agenda while OpenAI evaluated multiple sites.
Deloitte warns of a multi-year memory chip crunch reshaping server economics
A new Deloitte analysis said AI server DRAM prices roughly doubled in the first quarter of 2026 and could quadruple for the full year, with meaningful new capacity unlikely to come online before 2029 or 2030 as memory makers reallocate standard DRAM and NAND production toward high-bandwidth memory and enterprise SSDs. Hyperscaler capital spending is now projected to exceed $1 trillion in 2026, more than double the figure those companies planned in January, with memory absorbing roughly 30% of 2026 data center investment and a projected 36% in 2027. Memory now accounts for about a quarter of AI server rack bill of materials, and Deloitte forecasts memory sales surpassing $1 trillion in 2027 versus $230 billion in 2025. Because servers represent about 60% of the estimated $38 billion cost of building a typical one-gigawatt AI facility, Deloitte said rising memory content is already pushing per-megawatt build costs higher and creating longer wait times across hyperscalers, integrated server and storage OEMs, neoclouds, and both AI and non-AI operators, with consumers likely to feel secondary effects through PCs, smartphones, and telecom equipment.
Share this article
